ProcessGroupStream.h 8.1 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56
// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

#pragma once

#include "paddle/fluid/distributed/collective/ProcessGroup.h"

namespace paddle {
namespace distributed {

// NOTE(liyurui): Notice that some backends use `stream` as an abstract
// conception of hardward resource. We provide this base class allowing users to
// put communications on calculation stream. In some scenorios, we found this
// will save the time of switching streams.
class ProcessGroupStream : public ProcessGroup {
 public:
  class TaskStream : public ProcessGroup::Task {
   public:
    // TODO(liyurui): This constructor is temporary here for compatible reason,
    // will be deleted soon.
    TaskStream(int rank,
               const std::vector<phi::DenseTensor>& inputs,
               CommType comm_type)
        : Task(rank, inputs, comm_type) {}

    TaskStream(int rank,
               const std::vector<phi::DenseTensor>& inputs,
               CommType comm_type,
               bool sync_op,
               bool use_calc_stream)
        : Task(rank, inputs, comm_type, sync_op),
          use_calc_stream_(use_calc_stream) {}

    virtual ~TaskStream() = default;

   protected:
    bool UseCalcStream() const { return use_calc_stream_; }

   private:
    bool use_calc_stream_{false};
  };

  ProcessGroupStream(int rank, int size, const platform::Place& place, int gid);
  virtual ~ProcessGroupStream() = default;

57 58 59 60 61 62 63 64 65 66 67 68 69 70
  virtual phi::DeviceContext* GetDeviceContext(const Place& place,
                                               bool use_calc_stream) const;

  std::shared_ptr<ProcessGroup::Task> AllGather(
      std::vector<phi::DenseTensor>& in_tensors,   // NOLINT
      std::vector<phi::DenseTensor>& out_tensors,  // NOLINT
      bool sync_op) override;

  virtual std::shared_ptr<ProcessGroup::Task> AllGather(
      std::vector<phi::DenseTensor>& in_tensors,   // NOLINT
      std::vector<phi::DenseTensor>& out_tensors,  // NOLINT
      bool sync_op,
      bool use_calc_stream);

71 72 73 74 75 76 77 78 79 80 81 82
  std::shared_ptr<ProcessGroup::Task> AllReduce(
      std::vector<phi::DenseTensor>& input_tensors,   // NOLINT
      std::vector<phi::DenseTensor>& output_tensors,  // NOLINT
      const AllreduceOptions& options,
      bool sync_op) override;

  virtual std::shared_ptr<ProcessGroup::Task> AllReduce(
      std::vector<phi::DenseTensor>& input_tensors,   // NOLINT
      std::vector<phi::DenseTensor>& output_tensors,  // NOLINT
      const AllreduceOptions& options,
      bool sync_op,
      bool use_calc_stream);
83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227

  std::shared_ptr<ProcessGroup::Task> AllToAll(
      std::vector<phi::DenseTensor>& in_tensors,   // NOLINT
      std::vector<phi::DenseTensor>& out_tensors,  // NOLINT
      bool sync_op) override;

  virtual std::shared_ptr<ProcessGroup::Task> AllToAll(
      std::vector<phi::DenseTensor>& in_tensors,   // NOLINT
      std::vector<phi::DenseTensor>& out_tensors,  // NOLINT
      bool sync_op,
      bool use_calc_stream);

  std::shared_ptr<ProcessGroup::Task> AllToAllSingle(
      std::vector<phi::DenseTensor>& in_tensors,   // NOLINT
      std::vector<phi::DenseTensor>& out_tensors,  // NOLINT
      std::vector<int64_t>& in_sizes,              // NOLINT
      std::vector<int64_t>& out_sizes,             // NOLINT
      bool sync_op) override;

  virtual std::shared_ptr<ProcessGroup::Task> AllToAllSingle(
      std::vector<phi::DenseTensor>& in_tensors,   // NOLINT
      std::vector<phi::DenseTensor>& out_tensors,  // NOLINT
      std::vector<int64_t>& in_sizes,              // NOLINT
      std::vector<int64_t>& out_sizes,             // NOLINT
      bool sync_op,
      bool use_calc_stream);

  std::shared_ptr<ProcessGroup::Task> Broadcast(
      std::vector<phi::DenseTensor>& in_tensors,   // NOLINT
      std::vector<phi::DenseTensor>& out_tensors,  // NOLINT
      const BroadcastOptions& opts,
      bool sync_op) override;

  virtual std::shared_ptr<ProcessGroup::Task> Broadcast(
      std::vector<phi::DenseTensor>& in_tensors,   // NOLINT
      std::vector<phi::DenseTensor>& out_tensors,  // NOLINT
      const BroadcastOptions& opts,
      bool sync_op,
      bool use_calc_stream);

  std::shared_ptr<ProcessGroup::Task> Reduce(
      std::vector<phi::DenseTensor>& in_tensors,   // NOLINT
      std::vector<phi::DenseTensor>& out_tensors,  // NOLINT
      const ReduceOptions& opts,
      bool sync_op) override;

  virtual std::shared_ptr<ProcessGroup::Task> Reduce(
      std::vector<phi::DenseTensor>& in_tensors,   // NOLINT
      std::vector<phi::DenseTensor>& out_tensors,  // NOLINT
      const ReduceOptions& opts,
      bool sync_op,
      bool use_calc_stream);

  std::shared_ptr<ProcessGroup::Task> ReduceScatter(
      std::vector<phi::DenseTensor>& in_tensors,   // NOLINT
      std::vector<phi::DenseTensor>& out_tensors,  // NOLINT
      const ReduceScatterOptions& opts,
      bool sync_op) override;

  virtual std::shared_ptr<ProcessGroup::Task> ReduceScatter(
      std::vector<phi::DenseTensor>& in_tensors,   // NOLINT
      std::vector<phi::DenseTensor>& out_tensors,  // NOLINT
      const ReduceScatterOptions& opts,
      bool sync_op,
      bool use_calc_stream);

  std::shared_ptr<ProcessGroup::Task> Scatter(
      std::vector<phi::DenseTensor>& in_tensors,   // NOLINT
      std::vector<phi::DenseTensor>& out_tensors,  // NOLINT
      const ScatterOptions& opts,
      bool sync_op) override;

  virtual std::shared_ptr<ProcessGroup::Task> Scatter(
      std::vector<phi::DenseTensor>& in_tensors,   // NOLINT
      std::vector<phi::DenseTensor>& out_tensors,  // NOLINT
      const ScatterOptions& opts,
      bool sync_op,
      bool use_calc_stream);

  std::shared_ptr<ProcessGroup::Task> Send(
      std::vector<phi::DenseTensor>& tensors,  // NOLINT
      int dst_rank,
      bool sync_op) override;

  virtual std::shared_ptr<ProcessGroup::Task> Send(
      std::vector<phi::DenseTensor>& tensors,  // NOLINT
      int dst_rank,
      bool sync_op,
      bool use_calc_stream);

  std::shared_ptr<ProcessGroup::Task> Send_Partial(
      phi::DenseTensor& tensors,  // NOLINT
      int dst_rank,
      int64_t offset,
      int64_t length,
      bool sync_op) override;

  virtual std::shared_ptr<ProcessGroup::Task> Send_Partial(
      phi::DenseTensor& tensors,  // NOLINT
      int dst_rank,
      int64_t offset,
      int64_t length,
      bool sync_op,
      bool use_calc_stream);

  std::shared_ptr<ProcessGroup::Task> Recv(
      std::vector<phi::DenseTensor>& tensors,  // NOLINT
      int src_rank,
      bool sync_op) override;

  virtual std::shared_ptr<ProcessGroup::Task> Recv(
      std::vector<phi::DenseTensor>& tensors,  // NOLINT
      int src_rank,
      bool sync_op,
      bool use_calc_stream);

  std::shared_ptr<ProcessGroup::Task> Recv_Partial(
      phi::DenseTensor& tensors,  // NOLINT
      int src_rank,
      int64_t offset,
      int64_t length,
      bool sync_op) override;

  virtual std::shared_ptr<ProcessGroup::Task> Recv_Partial(
      phi::DenseTensor& tensors,  // NOLINT
      int src_rank,
      int64_t offset,
      int64_t length,
      bool sync_op,
      bool use_calc_stream);

  std::shared_ptr<ProcessGroup::Task> AllGather_Partial(
      std::vector<phi::DenseTensor>& in_tensors,
      std::vector<phi::DenseTensor>& out_tensors,
      int64_t offset,
      int64_t length,
      bool sync_op) override;

  virtual std::shared_ptr<ProcessGroup::Task> AllGather_Partial(
      std::vector<phi::DenseTensor>& in_tensors,   // NOLINT
      std::vector<phi::DenseTensor>& out_tensors,  // NOLINT
      int64_t offset,
      int64_t length,
      bool sync_op,
      bool use_calc_stream);
228 229 230 231
};

}  // namespace distributed
}  // namespace paddle